ReproductionRe: Section 4.1

Reproduction note: training seems sensitive to seed

TBtbecker· Independent· 6 days ago

Reproducing on CIFAR-10, I see noticeable FID variance (~1.5 points) across seeds with otherwise identical configs. EMA helps but doesn't fully remove it. Curious whether others observe similar seed sensitivity, and whether the variance shrinks with longer training.

Computer Vision ML Systems

2 Replies

YUyukis6 days ago

I saw the same on CIFAR-10. EMA decay of 0.9999 and 800k+ steps tightened it for me. Reporting mean±std over 3 seeds seems necessary for fair comparison.

TBtbecker6 days ago

Thanks — I'll extend training and report std. Marking this as reproduced with the seed caveat.